Artificial intelligence-enabled non-invasive ubiquitous anemia screening: The HEMO-AI pilot study on pediatric population

Abstract

Objective: Determine whether data collected from a smartphone camera can be used to detect anemia in a pediatric population.

Methods: HEMO-AI (Hemoglobin Easy Measurement by Optical Artificial Intelligence), a clinical study carried out from December 2020 to February 2023, recruited patients from the Pediatric Emergency Department, Pediatric Inpatient Department and Pediatric Hematology Unit of the Haemek Medical Center, Afula, Israel. A population-based sample of 823 patients aged 6 months to 18 years who had undergone a venous blood draw for a complete blood count since being admitted to the hospital were enrolled. Patients with total leukonychia, nailbed darkening or discoloration due to medication, nail clubbing, clinically indicated jaundice, subungual hematoma, nailbed lacerations, avulsion injuries, or nail polish applied on fingernails were not eligible for study recruitment. Video and images of the patients' hand placed in a collection chamber were collected using a smartphone camera.

Results: 823 samples, 531 from a 12.2 megapixel camera and 256 from a 12.2 megapixel camera, were collected. 26 samples were excluded by the study coordinator for irregularities. 97% of fingernails and 68% of skin samples were successfully identified by a post-trained machine learning model. Separate models built to detect anemia using images taken from the Pixel 3 had an average precision of 0.64 and an average recall of 0.4, whereas models built using the Pixel 6 had an average precision of 0.8 and an average recall of 0.84. Further supplementation of training data with synthetic data boosted the precision of the latter to 0.84 and the average recall to 0.87.

Conclusions: This study lays the groundwork for the future evolution of non-invasive, pain-free, and accessible anemia screening tools tailored specifically for pediatric patients. It identifies important sample collection parameters and design, provides critical algorithms for the pre-processing of fingernail data, and reports an initial capability to detect anemia with 87% sensitivity and 84% specificity.

Trial registration: Prospectively registered on www.clinicaltrials.gov (Identifier: NCT04573244) on 15 September 2020, prior to subject recruitment.

Keywords: Telemedicine general; adolescent medicine; connected devices personalized medicine; pediatrics medicine; remote patient monitoring personalized medicine.

Figure 1.

Overview of data collection and processing pipeline. (a) Signal extraction occurs by collecting 5–15 s of video on a smartphone camera from the patient’s hand situated in a collection chamber. The video is then parsed to individual frames. (b) Individual frames are put through a set of pre-processing steps. The post-trained YOLO model detects a swatch of skin from the middle finger and the four fingernails from the index, middle, ring, and little fingers. The skin swatch is matched to its corresponding monk skin tone. The fingernail images undergo a quality scoring and resizing pipeline, at the end of which a composite image of the top 100 fingernail frames is generated. The gain from 15 different color channels is extracted. (c) Three distinct models were trained and evaluated. The upper panel depicts the development of a model using images collected with a GP3, the middle panel depicts the development of a model using images collected with a GP6, and the lower panel depicts the development of a model using images collected with a GP6, which was supplemented with SMOTE-generated images for the training of the model.

Figure 2.

Flow diagram of patients enrolled in the Hemoglobin Easy Measurement by Optical Artificial Intelligence (HEMO-AI) study.

Figure 3.

Automatic fingernail detection. (a) A set of representative training samples for the post-training of the YOLO fingernail and skin detection model. (b) A set of representative labeled samples tagged by the post-trained YOLO fingernail and skin detection model. (c) Confusion matrix for the fingernail, skin, and background detection process. (d) Precision recall curve for the model is shown. Precision represents the positive predictive value of the algorithm, whereas recall represents the sensitivity of the model.

Figure 4.

Anemia detection model performance. (a) Mean area under the receiver operating characteristic curve of 0.75 for the XGBoost model with 95% CIs representing the variation of 20 independent model runs.